Abstract:

The amount of available data is increasing rapidly in almost every domain of our life. Thus, data science and machine learning face the challenge of bringing the valuable and essential data to the surface. To approach this challenge, models that can value sets of data by identifying and leveraging its complex dependencies are becoming more and more important.

In this talk, I will discuss approaches for learning models for valuing sets of data and show how to use these models in summarization and recommendation applications. A crucial assumption for tractability of learning and the application of such models is that of submodularity. I will first present algorithms for learning mixture models of fixed submodular component functions from data for valuing summaries of image collections. Then, I will introduce a probabilistic parameterized submodular model for encouraging diversity of items in recommendation applications, which can be seen as a parameterized component function, and illustrate how this model can be efficiently estimated from data. This corresponds to learning a higher-order probabilistic model, i.e. a probabilistic model that includes potentials depending on many random variables. This model can be naturally extended to include both log-submodular and log-supermodular higher-order potentials. I will demonstrate the effectiveness of our approach in a large set of experiments, where our model allows reasoning about preferences over sets of items with complements and substitutes. Finally, I will discuss potential directions for future research, combining recent advances in deep learning with the tractability of submodular function optimization.